Tamar Christina <tamar.christ...@arm.com> writes:
> Hi All,
>
> It seems I forgot to check that the operation we're combing when masking the
> predicated together are actually predicates types.
>
> Without it we end up accidentally trying to combine a value and a mask.
>
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
>
> Ok for master?
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> PR tree-optimization/103741
> * tree-vect-stmts.c (vectorizable_operation): Check for boolean.
>
> gcc/testsuite/ChangeLog:
>
> PR tree-optimization/103741
> * gcc.target/aarch64/pr103741.c: New test.
>
> --- inline copy of patch --
> diff --git a/gcc/testsuite/gcc.target/aarch64/pr103741.c
> b/gcc/testsuite/gcc.target/aarch64/pr103741.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..ef3ae66ebe5e5a44e7bea7be22b6378bc23cc538
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/pr103741.c
> @@ -0,0 +1,26 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-march=armv8-a+sve -O1" } */
> +
> +long int m, n;
> +
> +int
> +qux (int z)
> +{
> + return 4 >> z ? z : 0;
> +}
> +
> +int
> +bar (long int y)
> +{
> + return y ? 3 : 2;
> +}
> +
> +__attribute__ ((simd)) int
> +foo (int x)
> +{
> + long int a = x & m;
> + int b = bar (x) / n;
> +
> + return qux (b) == a;
> +}
> +
> diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
> index
> 8c427174b37e6c03c2f914c90332bcc4eac54130..ad90cdb0473a337207d6ba54c1dd0a2ecc50ab8d
> 100644
> --- a/gcc/tree-vect-stmts.c
> +++ b/gcc/tree-vect-stmts.c
> @@ -6361,7 +6361,9 @@ vectorizable_operation (vec_info *vinfo,
> /* When combining two masks check if either of them is elsewhere
> combined with a loop mask, if that's the case we can mark that the
> new combined mask doesn't need to be combined with a loop mask. */
> - if (masked_loop_p && code == BIT_AND_EXPR)
> + if (masked_loop_p
> + && code == BIT_AND_EXPR
> + && VECTOR_BOOLEAN_TYPE_P (vectype))
> {
> if (loop_vinfo->scalar_cond_masked_set.contains ({ op0,
> ncopies}))
Ah, so the reason that scalar_cond_masked_set.contains is not robust
against non-boolean op0 (despite looking like it might be) is that
we previously entered:
cmp = X != 0
into the set as <X, !=, 0>. Then looking up X implicitly becomes
<X, !=, 0> too.
So yeah, the patch is OK, thanks.
Richard
